Notes: Verify AI Results and Set Limits (Hallucinations and Guardrails)

Section 7 uses the worked comparisons to develop one verification ability: identify which claim or action could change the decision, choose evidence that can answer it, locate the failed move, and place control before consequences.

Lecture	Understanding developed
Why Does AI Sound Confident When It's Wrong? (Hallucinations)	Fluent language, a citation and evidence supporting the specific claim are separate. A model's confidence cannot create a missing publication record.
How Do You Check a Number in an AI Answer? (Fact-Checking)	A dated Trello answer is checked as a complete claim: product, plan, currency, paid-account count, billing period and arithmetic.
Which AI Claims Should You Verify First? (Risk and Evidence)	Daniel verifies the client-access requirement before cost because failed privacy can eliminate a product. The same priority can be applied to the demonstrated comparison or an optional personal result.
How Do You Find Where an AI Agent Went Wrong? (Look, Think, Do, Check)	Look, Think, Do and Check trace one hypothetical wrong total from source and rate selection to the delivered result and final verification.
Which Agent Actions Need Approval? (Guardrails and Human in the Loop)	Guardrails restrict permitted behaviour, a human in the loop reviews consequential actions, and an escalation path carries an unresolved case to a person.
What Do You Give an AI Agent Access To? (Permissions and Scopes)	Permissions and scopes are matched to the calendar actions and information Alex's task requires, including the separate effects of revoking access and deleting stored copies.
Can an Email Redirect Your AI Agent? (Prompt Injection)	An instruction inside an email remains untrusted content. The payment record grounds invoice status, and outbound replies remain behind approval.
What Can an AI Support System Miss? (Checking the Original Records)	Comparing an original record with a derived result can expose missing messages or facts. The worked example makes the comparison visible; a learner with a personal run can also inspect its original inputs.
Which Tasks Suit AI Agents? (Reliability and Human Checks)	Agents fit reviewable work with stated criteria and reachable evidence. The worked explanation connects correction with the approval boundary before any consequential action.

The Air Canada case remains dated and case-specific. The Trello answer is authentic saved data from 8 September 2026, while its screen recording and current-page check remain proposed. The support-queue defect is historical owner evidence.